Traffic Sign Recognition Based on ResNet-20 and Deep Mutual Learning

Traffic sign recognition has strong practical significance, such as driving assistance systems, urban planning and construction, automatic driving and even traffic monitoring content understanding. The traditional traffic sign recognition algorithm has the problems of low recognition accuracy and low robustness. In order to improve the classification accuracy and the self-adaptability of the network, this paper transfers the Deep Mutual Learning algorithm to the problem of traffic sign recognition. During the training process of the model, we use two ResNet-20 models to learn from each other. By training in this way, we find that the performance of each sub-network is better than when trained separately. Finally, we got two neural network models with higher robustness. In the prediction stage, we used the method of model fusion and obtained higher prediction accuracy. Through the verification of the test dataset, when only one model is used, Top-1 accuracy can reach 99.390% After using the method of Deep Mutual Learning, the Top-1 accuracy of model one is 99.462% and Top-S accuracy is 99.905%. The Top-1 accuracy of model two can reach 99.422% and the Top-5 accuracy is 99.881%, above the human performance of 98.81%. After using the method of model fusion, the Top-1 accuracy can reach 99.612% and Top-5 accuracy is 99.952% And we do the comparative experiment of this article and the recent mainstream methods. It shows that this algorithm has higher classification accuracy and stronger generalization ability.